ivasilyev1 commented on PR #59221: URL: https://github.com/apache/spark/pull/59221#issuecomment-5969242200
> > **AI-assisted review disclaimer:** This review was produced by an LLM-based coding agent. The > > reproducer and conclusions below were verified by executing the focused test against the stated > > PR commit. > > I think there is still a recovery hole for a checkpoint that has already gone through one failed restart on an affected Spark version. > > The sequence is: > > 1. Batch `N` is committed with five-partition state, but its V2 offset metadata has no > `spark.sql.shuffle.partitions` entry. > 2. The query is restarted with `spark.sql.shuffle.partitions=10`. Before loading state, the old > code writes offset `N + 1` with `10` copied into its metadata. > 3. Loading the five-partition state fails, leaving offset `N + 1` uncommitted but present. > 4. On the next restart, this PR sees that the latest offset contains the key and skips recovery. > It therefore tries ten partitions again and fails in exactly the same way. > 5. There should also be another restart after successfully re-executing `N + 1`: re-execution > commits the existing offset rather than rewriting it, so recovery must not leave the checkpoint > poisoned for the following restart. > ... @HeartSaVioR Yes agreed, this is the scenario I mentioned yesterday on the JIRA ticket but my comment didn't describe the issue as well as this -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
